BufferBenchmark.java

package com.googlecode.concurrentlinkedhashmap.benchmark;

import com.google.caliper.Runner;
import com.google.caliper.SimpleBenchmark;

import org.testng.annotations.Parameters;
import org.testng.annotations.Test;

import java.util.Queue;
import java.util.concurrent.ConcurrentLinkedQueue;

/**
 * This benchmark evaluates single-threaded performance of the buffer
 * approaches.
 *
 * @author [email protected] (Ben Manes)
 */
public class BufferBenchmark extends SimpleBenchmark {
  Queue<Integer> clq;
  MPSCQueue<Integer> mpsc;
  LockFreeQueue<Integer> lfq;

  @Override
  protected void setUp() {
    mpsc = new MPSCQueue<Integer>();
    lfq = new LockFreeQueue<Integer>();
    clq = new ConcurrentLinkedQueue<Integer>();
  }

  public int timeMPSCQueue(final int reps) {
    int dummy = 0;
    while (dummy < reps) {
      mpsc.enqueue(dummy);
      mpsc.dequeue();
      dummy++;
    }
    return dummy;
  }

  public int timeLockFreeQueue(final int reps) {
    int dummy = 0;
    while (dummy < reps) {
      lfq.enq(dummy);
      lfq.deq();
      dummy++;
    }
    return dummy;
  }

  public int timeConcurrentLinkedQueue(final int reps) {
    int dummy = 0;
    while (dummy < reps) {
      clq.add(dummy);
      clq.poll();
      dummy++;
    }
    return dummy;
  }

  @Test(groups = "caliper")
  @Parameters({"warmupMillis", "runMillis", "timeUnit"})
  public static void benchmark(String warmupMillis, String runMillis, String timeUnit) {
    String[] args = {
      "--warmupMillis", warmupMillis,
      "--runMillis", runMillis,
      "--timeUnit", timeUnit
    };
    Runner.main(BufferBenchmark.class, args);
  }
}

gistfile1.txt

Execution with params:
 - warmupMillis: 10000
 - runMillis: 5000
 - timeUnit: ns

[Caliper Benchmark]
Running BufferBenchmark
 0% Scenario{vm=java, trial=0, benchmark=MPSCQueue} 436.18 ns; ?=1.74 ns @ 3 trials
33% Scenario{vm=java, trial=0, benchmark=LockFreeQueue} 702.80 ns; ?=9.64 ns @ 10 trials
67% Scenario{vm=java, trial=0, benchmark=ConcurrentLinkedQueue} 34.99 ns; ?=0.06 ns @ 3 trials

            benchmark    ns linear runtime
            MPSCQueue 436.2 ==================
        LockFreeQueue 702.8 ==============================
ConcurrentLinkedQueue  35.0 =

LockFreeQueue.java

package com.googlecode.concurrentlinkedhashmap.benchmark;

import java.util.concurrent.atomic.AtomicReference;

/**
 * Lock-free queue.
 * Based on Michael and Scott http://doi.acm.org/10.1145/248052.248106
 * @author Maurice Herlihy
 */
public class LockFreeQueue<T> {
  private final AtomicReference<Node> head;
  private final AtomicReference<Node> tail;

  /**
   * Create a new object of this class.
   */
  public LockFreeQueue() {
    Node sentinel = new Node(null);
    head = new AtomicReference<Node>(sentinel);
    tail = new AtomicReference<Node>(sentinel);
  }
  /**
   * Enqueue an item.
   * @param value Item to enqueue.
   */
  public void enq(T value) {
    // try to allocate new node from local pool
    Node node = new Node(value);
    while (true) {                 // keep trying
      Node last = tail.get();    // read tail
      Node next = last.next.get(); // read next
      // are they consistent?
      if (last == tail.get()) {
        if (next == null) {        // was tail the last node?
          // try to link node to end of list
          if (last.next.compareAndSet(next, node)) {
            // enq done, try to advance tail
            tail.compareAndSet(last, node);
            return;
          }
        } else {                // tail was not the last node
          // try to swing tail to next node
          tail.compareAndSet(last, next);
        }
      }
    }
  }
  /**
   * Dequeue an item.
   * @throws queue.EmptyException The queue is empty.
   * @return Item at the head of the queue.
   */
  public T deq() {
    while (true) {
      Node first = head.get();
      Node last = tail.get();
      Node next = first.next.get();
      // are they consistent?
      if (first == head.get()) {
        if (first == last) {        // is queue empty or tail falling behind?
          if (next == null) {        // is queue empty?
            return null;
          }
          // tail is behind, try to advance
          tail.compareAndSet(last, next);
        } else {
          T value = next.value; // read value before dequeuing
          if (head.compareAndSet(first, next)) {
            return value;
          }
        }
      }
    }
  }

  /**
   * Items are kept in a list of nodes.
   */
  public class Node {
    /**
     * Item kept by this node.
     */
    public T value;
    /**
     * Next node in the queue.
     */
    public AtomicReference<Node> next;

    /**
     * Create a new node.
     */
    public Node(T value) {
      this.next  = new AtomicReference<Node>(null);
      this.value = value;
    }
  }
}

MPSCQueue.java

package com.googlecode.concurrentlinkedhashmap.benchmark;

import java.util.concurrent.atomic.AtomicReference;

class MPSCQueue<E> {
  Node<E> tail = new Node<E>(null);
  final AtomicReference<Node<E>> head = new AtomicReference<Node<E>>(tail);

  void enqueue(E e) {
    Node<E> node = new Node<E>(e);
    head.getAndSet(node).lazySet(node);
  }

  E dequeue() {
    for (;;) {
      Node<E> next = tail.get();
      if (next != null) {
        tail = next;
        return next.value;
      }
    }
  }

  static class Node<T> extends AtomicReference<Node<T>> {
    private static final long serialVersionUID = 1L;

    final T value;

    Node(T value) {
      this.value = value;
    }
  }
}

I wouldn't believe benchmarking frameworks so blindly...
Here is simple test and output that prove better performance of MPSCQueue:

private static final int ITERATIONS = 100000000;

public static void main(String[] args) {
BufferBenchmark bufferBenchmark = new BufferBenchmark();
bufferBenchmark.setUp();
for (int i = 0; i < 3; i++) {
final long t = System.nanoTime();
bufferBenchmark.timeMPSCQueue(ITERATIONS);
final long d = System.nanoTime() - t;
System.out.printf("MPSCQueue[%d]: %,d ns/op\n", i, d / ITERATIONS);
}
for (int i = 0; i < 3; i++) {
final long t = System.nanoTime();
bufferBenchmark.timeConcurrentLinkedQueue(ITERATIONS);
final long d = System.nanoTime() - t;
System.out.printf("ConcurrentLinkedQueue[%d]: %,d ns/op\n", i, d / ITERATIONS);
}
}

MPSCQueue[0]: 18 ns/op
MPSCQueue[1]: 16 ns/op
MPSCQueue[2]: 14 ns/op
ConcurrentLinkedQueue[0]: 35 ns/op
ConcurrentLinkedQueue[1]: 35 ns/op
ConcurrentLinkedQueue[2]: 35 ns/op